Does any data leave our network?
No. Intraplex runs entirely on your infrastructure. Documents, embeddings, indices, audit logs, and generated content stay local. Outbound network access is disabled by default. Air-gapped deployments stay current via signed offline update bundles.
Will this work in air-gapped and highly regulated environments?
Yes — that is the design point. SSO/RBAC, clearance-aware retrieval, immutable audit log, retention enforcement, and sensitive-data classification are part of the platform, not modules. Audit and classification logs export to CSV/JSON for SOC 2, GDPR, HIPAA, and FedRAMP programs.
What can it connect to, and do you respect permissions?
Read-only connectors for filesystems, S3-compatible storage, SharePoint, Google Drive, and major SQL and document databases (PostgreSQL, MySQL, SQL Server, Oracle, Snowflake, SQLite, MongoDB, Elasticsearch, ClickHouse, Redis, Neo4j). All connectors honor source-system ACLs. Permission trimming runs at retrieval, not just at display.
How does it handle scale and messy documents?
Production-validated at 50,000+ mixed-quality documents, with a roadmap to 1M+ via sharding and tiered storage. Document quality detection routes scans, OCR-heavy files, and native digital documents through appropriate pipelines. Vision models read tables, charts, and technical diagrams. Dense + sparse hybrid retrieval (SPLADE++) handles jargon-heavy corpora.
Which AI models does it use?
High-performance local models (Qwen, DeepSeek, Llama, GPT-OSS) for reasoning, Q&A, and research, with Anthropic and OpenAI-compatible providers available where policy permits. Multi-model orchestration is built in: per-team policies, per-project overrides, and task routing by complexity. Version pinning and rollback included.
What does deployment look like?
Single RTX-class server through multi-GPU clusters. CPU-only modes available for indexing or low-duty inference. Docker-based, with health checks and observability hooks. Works identically online or fully offline. Backup and restore are first-class.
How is sensitive data handled?
A background classifier scans every conversation for PII, PCI, PHI, Financial, CUI, Credentials, and PSI signals, assigning a risk level and confidence. The classification dashboard shows trends, top categories, and top users. Categories are configurable per organization. Classification runs async; it never blocks the chat.
What happens when someone asks for access to a document they're not cleared for?
The privileged-access agent acknowledges that relevant documents exist without exposing content. The user can submit a request with a justification; an admin approves with an optional time-limited grant (7 / 30 / 90 days or permanent) and a reviewer note. Grants expire automatically and can be revoked.
Does it produce real deliverables or just chat answers?
Real deliverables. The agent writes and executes Python in a sandbox to produce Excel models, PDFs, DOCX briefings, and charts — branded to your organization's templates. Every deliverable carries citations.
How is cost predictable?
Local models price like infrastructure, not like API usage. There are no per-token surprise invoices, no telemetry-based billing, and no upstream model deprecations forcing migration cycles. License tiers are flat; support is a percentage of license; custom work is scoped per engagement.